Key biomechanical features of jump-landing under cognitive dual-task conditions: an XGBoost-SHAP-based explainable machine learning analysis.

Objective: This study aims to quantify jump-landing characteristics and compare the performance of four machine learning (ML) models in classifying single- and dual-task conditions. Shapley Additive Explanations (SHAP) was used to identify key biomechanical features distinguishing the conditions and characterize multivariable biomechanical patterns associated with cognitive-motor interference, providing a quantitative basis for understanding associations between cognitive load and jump-landing biomechanics. Methods: Overall, 240 physically active young men completed single- and dual-task jump-landing tests. Kinematic and ground reaction force data were synchronously collected using a Qualisys high-speed motion capture system and an AMTI force plate. Hip, knee, ankle, and trunk angles were calculated using OpenSim inverse kinematics, and 13 kinematic and landing performance features were extracted. Four ML models were compared within the development set using participant-grouped nested cross-validation repeated five times. The selected model underwent held-out test-set evaluation; SHAP quantified feature contributions to its predictions. Results: XGBoost achieved the highest mean outer out-of-fold AUC in the development set (0.992 ± 0.002). Corrected paired comparisons supported higher AUCs for XGBoost than for the three comparator models (all Holm-adjusted P < 0.001). XGBoost's held-out test-set accuracy and AUC were 0.924 and 0.978, respectively. The five highest-contributing features were the normalized base-of-support ratio (BOS), hip flexion angle at initial contact (HF-IC), trunk flexion range of motion (TF-ROM), ankle dorsiflexion/plantarflexion angle at initial contact (PF-IC), and trunk flexion angle at initial contact (TF-IC). Under the dual-task condition, BOS and TF-ROM increased, whereas HF-IC, PF-IC, and TF-IC decreased, characterizing task-condition differences. Conclusion: The XGBoost-SHAP model identified and quantified key biomechanical changes during cognitive dual-task jump-landing, providing an explainable, quantitative description of dual-task landing biomechanics. These changes may be related to competition for limited attentional resources between motor control and cognitive tasks. The following five core features collectively characterize the biomechanical pattern distinguishing dual-task from single-task landing: increased BOS, decreased HF-IC, decreased PF-IC, increased TF-ROM, and decreased TF-IC. The model classified task condition only and was developed using data from physically active young men performing a laboratory mental-arithmetic task. Any injury-related interpretation requires prospective validation; generalization to other populations and task contexts requires external validation.